Capability · Framework — orchestration

Marvin

Marvin takes a 'Python-native' approach to LLM development: decorate a function with `@marvin.fn` and let an LLM implement it, or pass a Pydantic model to `marvin.extract` to pull typed data from text. It's designed to feel like adding AI to existing Python code rather than building a new application stack, which makes it a favourite for data pipelines and ETL-style tasks.

Framework facts

Category: orchestration
Language: Python
License: Apache 2.0
Repository: https://github.com/prefecthq/marvin

Install

pip install marvin

Quickstart

import marvin
from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int

people = marvin.extract(
    'Alice is 30 and Bob is 25',
    target=Person
)
# [Person(name='Alice', age=30), Person(name='Bob', age=25)]

Alternatives

Instructor — more popular Pydantic-based extraction
BAML — schema-first DSL alternative
Outlines — constrained generation with regex
Mirascope — decorator-based LLM library

Frequently asked questions

Marvin vs Instructor?

Instructor is more OpenAI-SDK-native and focused purely on structured output. Marvin is more opinionated and includes higher-level helpers (classify, cast, generate). Pick Instructor for minimal overhead, Marvin for more batteries-included data tasks.

Is Marvin still active?

Yes — it's maintained by Prefect and has seen steady 3.x releases with support for Anthropic, OpenAI, and custom providers.

Sources

Marvin — docs — accessed 2026-04-20
Marvin on GitHub — accessed 2026-04-20